DeepSeek

DeepSeek is the Chinese AI research laboratory, backed by quantitative trading firm High-Flyer, that shocked the global AI industry in January 2025 with the release of DeepSeek-R1 — an open-source reasoning model that matched or exceeded the performance of frontier models from OpenAI and Anthropic at a reported training cost of under $6 million, a fraction of the hundreds of millions invested by Western labs. The release triggered a $1 trillion market selloff in AI-adjacent stocks as investors re-evaluated the assumption that frontier AI requires unbounded capital expenditure.

The DeepSeek Shock

DeepSeek-R1's significance wasn't just technical; it was economic and geopolitical. The model demonstrated that innovative training techniques — particularly reinforcement learning applied to chain-of-thought reasoning — could achieve frontier performance without the latest NVIDIA hardware or hyperscale compute budgets. This challenged the prevailing assumption that AI leadership would be determined primarily by who spent the most on compute, and validated the thesis that algorithmic innovation can substitute for brute-force scaling. The reinforcement-learning method involved, GRPO, was introduced in DeepSeek's earlier DeepSeekMath paper and has since become a standard option in open training libraries. DeepSeek released its V4 family in April 2026, continuing to push the open-source frontier.

DeepSeek-V4 and Million-Token Context

DeepSeek's API changelog (opened 6 October 2026) records the sequence: V3.1 in August 2025, which merged thinking and non-thinking modes into one hybrid model; V3.2 in December 2025; and the V4 family on 24 April 2026. The V4 technical report (DeepSeek-AI, arXiv 2606.19348) describes two mixture-of-experts models — V4-Pro with 1.6 trillion parameters (49B activated) and V4-Flash with 284B (13B activated) — both with a one-million-token context window. Three changes carry the efficiency claim: a hybrid of Compressed Sparse Attention and Heavily Compressed Attention, Manifold-Constrained Hyper-Connections in place of plain residual connections, and the Muon optimizer, with pre-training on more than 32 trillion tokens. The report states that at one million tokens V4-Pro needs 27% of the single-token inference FLOPs and 10% of the KV cache of V3.2. Those are the developer's own measurements.

The line has moved quickly since. The changelog lists a V4-Flash official release on 31 July, a V4-Pro general-availability update on 13 August, an experimental vision variant on 21 August, and DeepSeek-V4.1-Flash on 10 September 2026. The V4.1-Flash model card describes a different design again: a 20-layer causal encoder feeding a 20-layer decoder, 552B backbone parameters with 8B active per token during prefill and 16B during decode, native image input, and a reasoning-effort setting expressed as an integer from 1 to 100. The V4 checkpoints opened on Hugging Face — V4-Pro-0813 and V4.1-Flash — carry the MIT licence, which places them among the most permissively licensed large open-weight models. Benchmark scores on the cards (for example 87.9 on Terminal Bench 2.1 for V4-Pro-0813) are vendor-reported.

API Pricing as of October 2026

DeepSeek's pricing page listed two hosted models on 6 October 2026, deepseek-flash (V4.1-Flash) and deepseek-v4-pro (V4-Pro-0813), each with a 1M-token context and a 384K maximum output. List prices ran from well under a dollar to a few dollars per million tokens, with off-peak rates at half the peak rates; the current figures are tracked on AI model pricing.

Context caching is on by default and works by prefix matching, so agent loops that resend a stable prefix pay the cache-hit rate on most of their input. The API is offered through both OpenAI-compatible and Anthropic-compatible interfaces, which lowers the cost of routing existing agent code to it. List price per token is not cost per task: reasoning models can spend many output tokens thinking, so comparisons belong at the task level.

Open Source and the Inference Economy

DeepSeek's commitment to open-source release accelerated the growth of the inference economy. Platforms like Groq and Together AI benefit directly from high-quality open-weight models that can be deployed on custom inference hardware, driving down the cost of AI deployment and enabling a wider ecosystem of applications. The open-source model also enables agentic engineering workflows where developers fine-tune and specialize models for specific tasks — a critical capability as AI agents become more specialized and domain-specific. In practice the V4 models are far larger than what most teams can fine-tune or self-host; their role is more often as a cheap hosted endpoint, a teacher for distillation, or a reference for techniques, while smaller open models serve as fine-tuning bases.

Geopolitics and AI Sovereignty

DeepSeek's emergence reshaped the geopolitical landscape of AI development. Despite U.S. export controls limiting China's access to advanced NVIDIA chips, DeepSeek demonstrated that China's AI ecosystem can produce competitive models through architectural innovation. This has implications for the global AI supply chain, for compute infrastructure investment patterns, and for the assumption that Western companies will maintain an insurmountable lead in foundation models. Combined with Alibaba's Qwen and other Chinese open-source models, DeepSeek is part of a growing multipolar AI landscape.